machine learning for noun phrase coreference · 2014-04-24 · last class ! noun phrase coreference...

Machine Learning for

Noun Phrase Coreference Claire Cardie

Department of Computer Science Cornell University

Last Class

§  noun phrase coreference resolution – what it is – why it’s important – why it’s hard

§  a (supervised) machine learning approach §  weakly supervised approaches

1.  Illustrate how much you’ve learned 2.  Realities of doing research in NLP+ML 3.  Introduce some cool weakly supervised learning methods

Noun Phrase Coreference

Identify all noun phrases that refer to the same entity

Queen Elizabeth set about transforming her husband,

King George VI, into a viable monarch. Logue,

a renowned speech therapist, was summoned to help

the King overcome his speech impediment...

Why It’s Hard

Many sources of information play a role –  string matching –  syntactic constraints – number agreement – gender agreement – discourse focus –  recency –  syntactic parallelism –  semantic class – world knowledge…

Why It’s Hard

§  No single source is a completely reliable indicator

§  Identifying each of these features automatically, accurately, and in context, is hard

Last Class

§  noun phrase coreference resolution §  a (supervised) machine learning approach

– evaluation – problems…some solutions

§  weakly supervised approaches

Knowledge-based approaches are still common. E.g.

- Lappin & Leass [1994]

- CogNIAC [Baldwin, 1996]

§  Classification – given a description of two noun phrases, NPi

and NPj, classify the pair as coreferent or not coreferent

[Queen Elizabeth] set about transforming [her] [husband], ...

coref ?

coref ?

coref ?

Aone & Bennett [1995]; Connolly et al. [1994]; McCarthy & Lehnert [1995]; Soon et al. [2001]; Ng & Cardie [2002]; …

A Machine Learning Approach

husband King George VI the King his

Clustering Algorithm

Queen Elizabeth her

Logue a renowned speech therapist

Queen Elizabeth

Logue

§  Clustering –  coordinates pairwise coreference decisions

[Queen Elizabeth],

set about transforming

[her]

[husband]

...

coref

not coref

not

coref

King George VI

A Machine Learning Approach

Machine Learning Issues §  Training data creation

§  Instance representation

§  Learning algorithm

§  Clustering algorithm

Supervised Inductive Learning

(novel) pair of NPs (features) class

Examples of NP pairs (features + class)

ML Algorithm

Concept description

(program)

Training Data Creation

§  Creating training instances –  texts annotated with coreference information –  one instance inst(NPi, NPj) for each ordered pair of NPs

»  NPi precedes NPj »  feature vector: describes the two NPs and context »  class value:

coref pairs on the same coreference chain not coref otherwise

anaphor candidate antecedent

Instance Representation §  25 features per instance

–  lexical (3) »  string matching for pronouns, proper names, common nouns

–  grammatical (18) »  pronoun_1, pronoun_2, demonstrative_2, indefinite_2, … »  number, gender, animacy »  appositive, predicate nominative »  binding constraints, simple contra-indexing constraints, … »  span, maximalnp, …

–  semantic (2) »  same WordNet class »  alias

–  positional (1) »  distance between the NPs in terms of # of sentences

–  knowledge-based (1) »  naïve pronoun resolution algorithm

Learning Algorithm

§  RIPPER (Cohen, 1995) C4.5 (Quinlan, 1994) –  rule learners

»  input: set of training instances » output: coreference classifier

§  Learned classifier »  input: test instance (represents pair of NPs) » output: classification

confidence of classification

Lie #1: Clustering Algorithm

§  Best-first single-link clustering – Mark each NPj as belonging to its own class:

NPj ∈ cj

– Proceed through the NPs in left-to-right order. » For each NP, NPj, create test instances, inst(NPi, NPj), for

all of its preceding NPs, NPi. » Select as the antecedent for NPj the highest-confidence

coreferent NP, NPi, according to the coreference classifier (or none if all have below .5 confidence);

Merge cj and cj .

Plan for the Talk

§  noun phrase coreference resolution §  a (supervised) machine learning approach

– evaluation – problems…some solutions

§  weakly supervised approaches

Evaluation

§  MUC-6 and MUC-7 coreference data sets §  documents annotated w.r.t. coreference §  30 + 30 training texts (dry run) §  30 + 20 test texts (formal evaluation) §  scoring program

–  recall –  precision –  F-measure: 2PR/(P+R)

System output

C D A B

Key

Baseline Results

MUC-6 MUC-7

R P F R P F Baseline 40.7 73.5 52.4 27.2 86.3 41.3

Worst MUC System 36 44 40 52.5 21.4 30.4

Best MUC System 59 72 65 56.1 68.8 61.8

machine learning for noun phrase coreference · 2014-04-24 · last class ! noun phrase coreference...

Documents